Machinery Fault Diagnosis Method Based On Dictionary Learning Theory

After the world's first Industrial Revolution,modern technology and industry has achieved great development.At the same time,there exists urgent demand for large-scale,integrated,high-speed and precise machinery equipment.The safe,stable,efficient operation of equipment ensures the economic benefits.Once they fail unexpectedly,it usually leads to severe economic loss,threatens people life safety and causes environment pollution.Therefore,condition monitoring,fault diagnosis and fault prevention is very important and meaningful in economic benefit and safety.Taking the key components in rotating machine as research object,this thesis mainly focused on the application of dictionary learning method on mechanical feature extraction and fault diagnosis.The researches of this thesis mainly focus on three aspects,i.e.,fault feature extraction based on normal dictionary learning;fault feature extraction based on dictionary learning with special structure;fault diagnosis based on dictionary database and sparse feature.The main contents are as follows:(1)From the perspective of engineering application and safety control,we discuss the significance of the selected topic.This thesis mainly summarizes several important issues,such as feature extraction methods,intelligent diagnosis methods,performance degradation assessment methods and residual life prediction.Besides,we also summarize the recent development of sparse representation and dictionary learning theory.Finally,the research contents are introduced.(2)The mathematic theory of sparse representation and dictionary learning are reviewed.We mainly introduce several important concepts such as over-complete representation,redundant dictionary and so on.Also,the related algorithms are also presented and discussed.This chapter lays a solid theoretical support for the following chapters.(3)The sparse representation model for noisy signal is introduced.And we present the de-noising method based on KSVD dictionary learning.By kinematic analysis of rolling element bearing,the main characteristic frequencies are derived.Then,a simulation bearing signal is used to validate the fault feature extraction based on KSVD.And the influences of main parameters on noise reduction are discussed.Finally,considering that KSVD is sensitive to noise,a method combined with minimum entropy deconvolution and KSVD is proposed in this paper.Experiment proves that this method can extract bearing incipient weak fault under strong background noise.(4)This thesis introduces a kind of dictionary with special structure – shift-invariant dictionary.The decomposition of circulant matrix is used to solve this shift-invariant dictionary learning(SIDL)problem.Then,we propose a fault feature extraction method based on shit-invariant dictionary learning in this thesis.The proposed method is very effective in extracting repeating features in signal.Through a simulation bearing signal,we present the procedure of the proposed method,and proves its validity.The two experiments,including bearing double-impulse structure and whole life test,proves the validity of the proposed method.Furthermore,comparisons with other methods are also made.(5)A single channel blind source separation based on shift-invariant correlation analysis is proposed.In order to separating multiple sources in single channel,the proposed method combines shift-invariant dictionary learning and dictionary correlation analysis in spectral domain.A new concept – between-class correlation coefficient is proposed.And the source number is determined by minimizing the mean between-class correlation.Through the simulation and experiments such as bearing and gearbox compound fault,the procedure of the proposed method is presented in detail.The experimental results show that the proposed methodology is feasible and effective.(6)By fusing all sub-dictionaries for each condition,a whole dictionary database is built.Based on this dictionary database,a sparse feature extraction method based on dictionary learning is proposed.Later,by combining a method combined with sparse feature and hidden Markov model is proposed,and its framework is presented in detail.Two experiments,including bearing single fault and compound fault experiment,are utilized to prove the validity and effectiveness of the proposed method.